We investigate on the temporal dynamics phenomenon in recommender systems. By analyzing the public dataset from real world applications, we find the temporal dynamics phenomenon is common in the online recommender systems, and the phenomenon would cause problems in making good recommendations. In this thesis, we propose four approaches to tackle the problems caused by the temporal dynamics phenomenon. The four approaches are the user's autoregressive interests evolution, user's markovian interests evolution, a POMDP recommendation framework, and the transfer learning approach. Both the user's autoregressive interests evolution and the user's markovian interests evolution are motivated by the sequential property in the changes of the user's interests. The POMDP recommendation framework is inspired by the self-learning mechanism of reinforcement learning models. The transfer learning approach is driven by the rich source domain data. Overall, the four approaches focus on handling the problems raised by temporal dynamics phenomenon in recommender systems. We also discuss the metrics and the datasets to verify our proposed approaches.